Reintroduction to Concurrency

Transcription

1 Reintroduction to Concurrency The execution of a concurrent program consists of multiple processes active at the same time. 9/25/14 7

2 Dining philosophers problem Each philosopher spends some time thinking and some time eating In order to eat, he needs two forks When he finishes thinking he picks up the forks on either side of him, eats for a while, then replaces the forks and starts thinking again How do we ensure that no philosopher starves? What are the properties of the system? 8

3 Exercise Solve the Dining Philosophers problem! 8 to a group 10 minutes max Devise a strategy that each philosopher can follow blindly (i.e. an algorithm) No philosopher should starve Test your strategy with your group No Solution? What are the problems? A Solution? Show us. 9

4 Philosophical Possibilities 1. Each philosopher simultaneously picks up the fork on his left, then goes for the one on his right, which will be held by another philosopher deadlock 2. A philosopher can simultaneously pick up the two forks when they become available atomic actions 3. An important philosopher can take a fork away from a less important one pre-emptive acquisition 10

5 Why do things concurrently? Performance execute as fast as possible Throughput get as much done as possible Efficiency minimise the time spent unproductively Programmability some problems are naturally concurrent what problems? 11

6 Logical concurrency Several happen together in the designer s mind Accept new orders Dispatch good from the factory ACME ACME ACME 12

7 Multitasking Multitasking is where a single processor executes more than a single activity Shares its times between the activities Periodically the processor changes from one task to another Tasks run together, rather than one after the other Executes a context switch to change its state The order and frequency of context switches are controlled by a scheduling algorithm 13

8 True concurrency True concurrency sometimes referred to as parallelism happens when a system has more than one processor working on a job Multiprocessor workstation Dedicated parallel computer Co-operating systems on a network 14

9 Processes Most operating systems have the notion of a process An executing program Typically, processes are somewhat isolated from each other each owns An address space A set of resources such as open files, sockets etc. By default, processes don t directly interact No interference No co-operation 15

10 Concurrency in processes A system will usually have several processes running The processes co-operate to provide a system, rather than in providing a single task Each task is provided by a single simple program of the normal kind a sequence of instructions 16

11 Threads Threads (sometimes called lightweight processes) Lots of executing sequences within a single process Share memory, files, windows, Can interact Can interfere Threads are provided by some operating systems as standard (Win32, Solaris), and by others as run-time libraries (some Unix variants) 17

12 Thread scheduling How do we order the execution of threads? Are all threads equally important? If yes give each the same time on the processor If no give some more time, or run them preferentially Can the programmer force a thread to keep running? If yes rely on the programmer to let other threads have a chance If no forcibly evict a thread and run another 18

13 Co-routines Co-routines are a co-operative approach to multitasking A thread starts running, and keeps running until it voluntarily gives up the processor First thread yields the processor to second thread Second thread yields, first thread resumes If a thread doesn t yield, it keeps the processor Locks out other threads Lets the programmer keep running something important More control, more work 19

14 Pre-emptive scheduling Simplicity doesn t offer enough compensation for the disadvantages of co-operative scheduling All too easy for a thread not to yield Can lead to some very convoluted programming structures Almost all modern systems use pre-emptive scheduling, where context switches occur based on the rules of the scheduling algorithm No lock-outs, fairer, more potential problems You still encounter co-routines in some embedded real-time systems where fine control is too important to lose 20

15 Interleaving If the processor s being shared, there s the possibility that actions from one task will be mixed-in with actions from another This is known as interleaving 21

16 Interleaving Interleaving also occurs in truly concurrent systems Any concurrent computation will behave, in the final analysis, like some interleaved sequence of its basic indivisible actions Interleaving of actions is the key concept in concurrent programming 22

17 Interleaving Any kind of interleaving has an important side effect Suppose the two tasks share some structure or resource. Interleaving their actions can cause some rather bizarre effects Open the file Read its contents Modify and re-write Close the file Open the file Read its contents Open the file Delete some records Modify and re-write Close the file Close the file Open the file Delete some records Close the file 23

18 Interference This phenomenon is referred to as interference the actions of two tasks, correct in themselves, conflict when combined concurrently Interference is the basic correctness problem in concurrent programming The essence of the problem Tasks use several actions to manipulate the shared structure and these actions are mixed up together and the effects of actions in one task conflicts with the behaviour of actions in the other 24

19 Interference inconsistency Some actions are naturally tied together people: totalmarks: newmark() add 1 to people add total mark re-compute average average() get number of people get total marks divide one by the other Interleaving made the structure inconsistent for a time in which period a calculation took place on bogus data

20 Interference race conditions The is an example of a race condition Different results depending on exactly which action runs first newmark() average() 60 newmark() average() 61 newm average() ark() 57 26

21 Atomicity Each routine is composed of several actions, and those actions must occur indivisibly atomically for the result to be correct This idea of atomicity is the key to avoiding undesirable artefacts from concurrency Identify the critical sections of code that must execute without any extraneous actions in between them Make these sections atomic 27

22 Locking and signalling To implement shared data structures without interference, we need to do one of two things Signal when a thread is using the structure, so other threads can check whether it s safe to modify it Lock the structure so that, if one thread is accessing it, no-one else can Signalling protocols are still quite co-operative and flexible, as a thread gets to check and choose whether to change a shared structure Locking is safer and more structured, but sometimes a little restrictive 28

23 Synchronisation Some concurrently executing tasks have a natural ordering with respect to each other Tasks need to be able to synchronise with each other to enforce such orderings 29

24 Semaphores A semaphore is an abstract data type with two operations, wait() and signal() wait() waits for the semaphore to be dropped, then raises it and continues in one atomic action signal() drops the semaphore Associate a semaphore with every shared structure and bracket all accesses with wait() signal() E.W. Dijkstra, Co-operating sequential processes, in Programming languages, ed. F. Genuys, Prentice-Hall (1968) 30

25 Monitors Monitors enforce mutual exclusivity on mutually interacting critical sections of code The actions in the monitor are the only ones able to manipulate the data structure This call will block until the first has completed Internally, of course, the monitor is just a safe wrapper round a semaphore C.A.R. Hoare, Monitors: an operating system structuring concept, 9/25/14 Communications 2004 Trinity College of the Dublin ACM 17(10) (October 1974) 31

26 Monitors Monitors can provide language-level concurrency control The language construct can be checked and analysed Programmers can t avoid the concurrency control, accidentally or maliciously Less error prone when compared to sockets Monitors are the basis of concurrency control in lots of modern programming languages 32

27 Monitors and objects In object-oriented languages monitors are typically associated with objects Object = data + operations so if we wrap a monitor around the operations, we protect the data from corruption This notion of a protected value is very useful Typically speak of locking a variable while it s being accessed, and blocking on a variable waiting for it to become free 33

28 Problems, problems computationone() lock a lock b add 10 to a + b store the result in b a: 22 b: 79 Blocks waiting for b Blocks waiting for a lock a lock b lock b lock a computationtwo() lock b lock a subtract 10 from b * a store the result in a 34

29 Deadlock This situation everyone waiting for everyone else, with no-one making progress is called deadlock Well studied problem in computer science Several solutions, but no really general ones circular waiting: each task is waiting for (one or more of) the others There are some very convoluted cases not always just two tasks It appears in applications with depressing regularity, and we ll look at it more later 35

30 Deadlock Deadlock may occur when code has the following properties: Piecemeal acquisition of resources Locks one at a time Resources acquired in any order Circular waiting Ability to hold lock indefinitely 36

31 Concept summary Activity Processes, threads Actions, interleaving Critical sections, atomicity Synchronisation Signalling and locking Constructs Semaphores Blocking Monitors, objects Deadlock 37

32 Java Threads In Java a process is represented by a thread. To make a thread run you call its start() method. This registers the thread with the thread scheduler. start() does not cause the thread to run immediately it only makes it eligible to run. The thread must contend with other threads for the CPU. When a thread gets to execute, it executes a run() method: 38

33 threads in Java A Thread class manages a single sequential thread of control. Threads may be created and deleted dynamically. Thread run() MyThread run() The Thread class executes instructions from its method run(). The actual code executed depends on the implementation provided for run() in a derived class. class MyThread extends Thread { public void run() { //... Thread x = new MyThread(); 39

34 threads in Java Since Java does not permit multiple inheritance, we often implement the run() method in a class not derived from Thread but from the interface Runnable. Runnable target Thread run() MyRun run() public interface Runnable { public abstract void run(); class MyRun implements Runnable{ public void run() { //... Thread x = new Thread(new MyRun()); 40

35 threads in Java Write a Java program that creates three Java threads (by extending Thread). The threads should sleep for a specified amount of time and then print their name and the length of time they slept for to the console. The three threads should sleep for 500ms, 250ms, 1000ms respectively. Repeat the exercise implementing Runnable 41

36 Extend Thread Example class ExtendThread extends Thread{ private int sleeptime; public ExtendThread(String name, int time){ super(name); sleeptime = time; public void run(){ while(true){ try{ sleep(sleeptime); catch (InterruptedException e){ System.out.println(getName() + " " + sleeptime); 42

37 Extend Thread Example class ExtendMain{ public static void main(string args[]){ ExtendThread t1 = new ExtendThread("Thread 1", 500); ExtendThread t2 = new ExtendThread("Thread 2", 250); ExtendThread t3 = new ExtendThread("Thread 3", 1000); t1.start(); t2.start(); t3.start(); 43

38 Implement Runnable Example class ImplementRunnable implements Runnable{ private Thread t; private int sleeptime; public ImplementRunnable(String name, int time){ sleeptime = time; t = new Thread(this, name); public void start(){ t.start(); public void run(){ while(true){ try{ Thread.sleep(sleepTime); catch (InterruptedException e){ System.out.println(t.getName() + " " + sleeptime); 44

39 Implement Runnable Example class ImplementMain{ public static void main(string args[]){ ImplementRunnable t1 = new ImplementRunnable("Thread 1", 500); ImplementRunnable t2 = new ImplementRunnable("Thread 2", 250); ImplementRunnable t3 = new ImplementRunnable("Thread 3", 1000); t1.start(); t2.start(); t3.start(); 45

40 Executing a Thread class Interleave { public static int c1 = 2; public static int c2 = 3; public static void main (String[] args) { Thread p1 = new P1 (); Thread p2 = new P2 (); p1.start (); p2.start (); Extend the Thread class and override its run() method. class P1 extends Thread { public void run () { Interleave.c1 = Interleave.c1 *Interleave.c2; class P2 extends Thread { public void run () { Interleave.c1 = Interleave.c1 +Interleave.c2; 46

41 Executing a Thread class Interleave { public static int c1 = 2; public static int c2 = 3; public static void main (String[] args) { Thread p1 = new Thread(new P1()); Thread p2 = new Thread(new P2()); p1.start (); p2.start (); Sometimes it is desirable to implement the run() method in a class not derived from Thread but from the interface Runnable. class P1 implements Runnable { public void run () { Interleave.c1 = Interleave.c1 *Interleave.c2; class P2 implements Runnable { public void run () { Interleave.c1 = Interleave.c1 +Interleave.c2; 47

42 Executing a Thread When the run() method ends the thread is considered dead. A dead thread cannot be started again, but it still exists and, like any other object, its methods and data can still be accessed. 48

43 Thread States When a thread s start() method is called, the thread goes into a ready-to-run state and stays there until the scheduler moves it to the running state. In the course of execution the thread may temporarily give up the CPU and enter some other state. Running Monitor states Suspended Asleep Blocked Ready 49

44 Thread states - Yielding If there are any other threads in the Ready state, the thread that just yielded may have to wait before it gets to execute again. If there are no waiting threads in the Ready state the thread that just yielded will get to continue executing immediately. Running Thread.yield() scheduled Ready 50

45 Thread states - Sleeping A call to the static sleep() method requests the currently executing thread to cease executing for an approximately specified period in milliseconds. Running Ready Asleep Note: when the thread finishes sleeping it does not continue execution directly. 51

46 Thread states - Blocking If a method needs to wait for an indeterminable amount of time until some I/O occurrence takes place it should step out of the Running state. This is know as blocking. All Java I/O methods behave this way. Running Ready Blocked A thread can also become blocked if it fails to acquire the lock for a monitor or if it issues a wait() call. This will be explained later. 52

47 Thread Priorities and Scheduling Every thread has a priority (from 1..10). All newly created threads have their priority set to that of the creating thread. Higher priority threads get preference over lower priority threads. The scheduler generally chooses the the highest-priority waiting thread. 53

48 Thread Priorities and Scheduling If there is more than one waiting thread the scheduler chooses one of them. There is no guarantee that the one chosen is the one that has been waiting longest. int oldpriority = athread.getpriority(); int newpriority = Math.min(oldPriority+1, Thread.MAX_PRIORITY); athread.setpriority(newpriority); Note: The way thread priorities affect scheduling is platform dependent. 54